I like that Agent 3 can carry out more tasks at once and do minor testing/checking as this saves me having to test (to discover what else broke). Agent 3 feels no smarter than the prior models in terms of capability, but because Agent 3 can work longer and analyse more code/do more work, Agent 3 gets the answer right quicker than prior models (where I as the user was involved more in testing). Agent 3 works harder, not smarter. The job gets done quicker. Tick.
Learning
My biggest learning with Agent 3 which I wish someone told me before I started using Agent 3 is that the types of changes it will make to your app are significantly different to the types of changes in prior models (because it works harder, not smarter), therefore you may get what you think is a win in terms of progressing your app, but hold your breath. Stop, publish/deploy the change to ensure nothing hidden in the background broke. I did not do this and ploughed ahead with heaps of brilliant changes to my app (replit team - check out how much revenue you got off me, it’s insane!!) However now I’ve achieved what I want I’m unable to publish any changes. I’ve spent over US$100 just trying to deploy app so that it runs fine in the live environment like 2 days ago, however no matter what way I ask Agent 3 to analyse and resolve deployment issues, and now matter how much money or time I throw at Agent 3, it’s as if Agent 3 goes around in circles trying fixes it already knows doesn’t work. Very frustrating. I had an app working, spent a fortune making incremental improvements, now have no app at all in the live environment.
Negative
Slower and less “smart” – Several felt Agent 3 performs worse than Agent 2 in speed, reasoning, and usefulness. (6+ users, Moderate concern). I feel Agent 3 is more convincing that it has e.g. correctly identified and fixed an issue, however I observe issue persists.
Auth testing bugs – Agent 3 overrides authentication, resets passwords, or fails to handle login flows correctly. (4+ users, Major concern). Yeah, this gave me a fright.
Fees: insanely insanely high relative to previous pricing. E.g. 12 hours of coding yesterday cost me more than my coding from inception (May 2025) through to yesterday. Because value (improvements to my app) for money has reduced by about 10 fold (my subjective sentiment, even after considering the huge value add of getting changes done to my app faster than what I reckon I could have done with prior model), I’m basically sidelined until Agent 4, which I hope will be smarter, and give me the option to turn it into: work smart mode, or work hard mode.
I want Replit to win, I back you guys, you’ll figure this out! Just make the AI smarter so I can deploy my app!!!
I used to be able to plan, put a list together of 5–10 changes on a page and it would get it done using the 1.25x cost we had before in about 5 minutes and with 95% accuracy for less than $2.
Now, using the same methods and using the plan tool and all the other features off, it works for 15 minutes, charges me $6, and is about 20% accurate.
Current Situation
I am posting on here out of sheer desperation before I pack up and leave Replit for a competitor.
PLEASE give us the option to use V2 or V3.
Allow us to rollback to whatever we were on a week ago so I can go back to my hyper-efficient Webflow I was working on.
I have built over 50 applications and websites with Replit since February and my productivity and output is meteoric, but I am literally dead in the water and have spent nearly $300 in prompts this week alone and I have nothing to show for it.
Literally trying any of my apps and building on them and I’m left with 100% rollbacks of which I’m still charged for.
This is a complete nightmare and it doesn’t take more than a 5-second Google search to see I’m not alone.
User Experience Failures
I just tried a prompt where I built a plan and it went to execute on it.
8 minutes in, it had made more of a mess, so I tried to hit stop.
It wouldn’t stop, I had to kill my browser and reboot just to get it to stop.
It’s like it’s broken loose of its parameters and has lost all benefit.
Broader Strategic Failure
AI is not smart enough to run successfully, bug-free, and autonomously for anything more than maybe 15 minutes at best.
The planning tool is great, but we should have more “modes” that we can use to function in different roles:
Planning
Testing
Building
Rapid mode for short quick fixes across the apps
There should be more transparency on charges:
You no longer see total charges per conversation, just per prompt.
There’s no total spend per app on prompt fees or any analytics to track the cost between different apps regarding prompt fees.
The ability to pay for smarter models was great with the 1.25x model.
The higher pricing model I found too expensive for the output, but the 1.25x was a great sweet spot.
Summary
I am taking the rest of the week off until I hear back on changes because I’ve dumped way too much money this week into Replit with nothing to show for it.
I love Replit and have spent $1000s with this company as a vibe coder and am building my own little empire off of it.
But I will be forced to move within the next 1–2 weeks at most if significant improvements in cost, speed, and results are not back in the agent.
Yes, i tried to update a landing page on a site i was on and after 15 minutes of work it had broken 4 things, removed 3 others not even on the list and i tried to rollback and it wouldnt stop coding. I tried hitting stop and it didnt stop it literally just kept coding and destroying. Once it was done it was a $6~ charge and 15 minutes of destruction. I rolledback only after restarting browser then it stopped and allowed rollback.
i’ve built so much on replit and it has destroyed more and cost more than i can even handle this week
Not to mention i have emailed support 3x times this week regrading these problems and getting 0 response back from the…. i dont know what happened to this company but it has been ruined, i really hope they make some massive changes to the response of these issues
I already rolled back so I do t see the charge. Hence part of The price transparency issue
I know app preview and max autonomy were turned off though as I disabled those Monday after the $150 of charges across the apps I was working on and I realized that was bloating time, costs and results.
sorry i replied on another account i have, we’re the same person though i’m not sure how to get this resolved, right now i’m self auditing very closely and slowing way way down. sure wish i could get back to the speed and workflow i had in V2, i’ll monitor and screenshot more errors today, maybe someone from replit will actually reply back with some solutions? let me know if you find any
I’m having the same issue. More then 3k into an app. It keeps not fixing the things I task it with and break things not even related. I have also taken a break as it’s costing too much just to get back to 0. Also the price indicated on the prompts don’t always match the total in the usage.
I’ve been trying to locate Extended Thinking Mode in my Replit UI but haven’t been able to find it anywhere. I reached out to support previously, and they mentioned it should appear in the toolbar. I even shared a screenshot showing that the option isn’t available on my end. At that point, I was told I’d receive further assistance, but I haven’t heard back yet.
I’d like to confirm whether Extended Thinking Mode is still an available feature, or if it may have been deprecated. I’m aware of High Power Mode, but I’ve used Extended Thinking extensively in the past (prior to Agent 3) and found it extremely useful.
Could you please clarify whether this feature is still supported, and if so, how I can access it?
i used agent 3 with high power mode for 30 minutes with medium level and it started working again, it made a few changes which were OK not phenomenal, it racked up $120 in charges.
it looks like they removed 1.25x extended thinking and just kept the high power model, which is insanely expensive to use. As of right now high power model is the only way my agent works, so I can’t use it, its charging me about $120 every 30 minutes to use it, i might as well hire actual developers at that point, in house to finish the project.
High power mode is very costly. It’s direct cost to replit is 500% more than regular modes, so you can expect quite literally 500% increase in your cost.
Enterprise-grade: RBAC, audit trails, medical data compliance
This is a substantial enterprise SaaS platform with sophisticated architecture, comprehensive feature set, and production-ready codebase suitable for multi-tenant system at scale.
Exactly the same issue. I was trying to fix a few “final touches” when v3 published, now spent 50usd and i cant even open the app, trying to fix an “internal server error” i didnt have before agent broke it. Unbelievably frustrating. Idk if anyone figured out how to work with v3. moneypit.
I just got charged $0.13 from this prompt "Give me the structure of the database of this app so I can create it as an SQL database elsewhere”. So yes, we are being charged for internal compute. Not for much longer, as I will be cancelling my subscription. All trust is gone and I don’t want to be wedded to these people:
i’m not sure how to get this resolved, right now i’m self auditing very closely and slowing way way down. sure wish i could get back to the speed and workflow i had in V2, i’ll monitor and screenshot more errors today, maybe someone from replit will actually reply back with some solutions? let me know if you find any
Overall Statistics
Architecture Breakdown
Key Insights